Architecture overview

ModelBump is a pipeline of independent stages. Each stage has one job, a typed input and a typed output, so stages can be tested, evaluated and replaced on their own.

flowchart LR
    code[Target repo:<br/>code, configs, tests] --> scan[1 Scanner]
    registry[(Retirement<br/>registry)] --> scan
    scan --> synth[2 Suite synthesizer]
    synth --> runner[3 Runner<br/>cache + budget]
    runner --> stats[4 Stats engine]
    stats --> repair[5 Prompt repair]
    repair --> runner
    stats --> report[6 Reporter]
    repair --> pr[7 GitHub PR]
    report --> pr

Stage	Responsibility	Status
Registry	Which model IDs are retired, and when	v1 (Phase 1)
Scanner	Find model IDs in code and config; match them against retirement dates	v0.1.0 (Phase 2)
Suite synthesizer	Build regression test cases and oracles from the target repo	Planned (Phase 5)
Runner	Call models with caching, retries and a cost budget	Done (Phase 3)
Stats engine	Decide REGRESSED / IMPROVED / EQUIVALENT / INCONCLUSIVE per test case	In progress (Phase 4)
Prompt repair	Find prompt edits that fix regressions safely	Planned (Phase 6)
Reporter / GitHub	Produce reports and open the upgrade PR	Planned (Phase 7)

Each stage gets its own page here when it is built.